Reasoning effort is a dial, not a vendor lock-in
Warden gives you one reasoning-effort control, off through max, that maps onto every provider it drives, shows you the thinking tokens it's spending, and respects your session cost cap. Set it per session, per command, or per agent.
Every frontier model now has some notion of “think harder,” whether it calls that extended thinking, reasoning effort, or a thinking budget. The problem is that each vendor exposes it differently, so the moment you use more than one provider, “think harder” becomes a per-vendor configuration puzzle. Most clients respond by picking one vendor and hard-wiring to it.
Warden does the opposite. It gives you a single dial (off, low, medium, high, max) and translates it onto whichever provider is actually running the task. One control, every model.
One ladder, every provider
You set the effort; Warden maps it natively. Ask for high effort and a model that supports extended thinking gets it; a model that exposes a coarser setting gets the closest honest equivalent. You don’t memorize each vendor’s knob, and you don’t lose the capability when you switch lanes mid-session. The dial means the same thing everywhere it can.
You can see what thinking costs
Reasoning isn’t free, and Warden refuses to hide the bill. The thinking tokens a step spends are surfaced right in the interface, in the footer as you work and in the startup banner, so “think harder” is a visible trade rather than a silent line item you discover at the end of the month. And it stays inside the guardrail: your session cost cap still governs, so cranking effort to max can’t quietly blow past the ceiling you set.
Set it where it belongs
Effort isn’t one global switch you flip and forget. Drive it for a whole session with a flag, change your mind mid-flight with a command, or pin a specific effort to a specific agent so a cheap mechanical task stays cheap and a hard one gets room to think. The control lives at the altitude the decision actually happens.
Where it stands
This is live in Warden today: pre-alpha, MIT, still no version promises. It pairs with TheAuditor for ground truth, Curator for memory, and Arbiter for orchestration across providers. The companion piece is the TUI that never truncates; the feed is the RSS.
Was this useful?